As someone who works in NLP, I tend to be skeptical of generative models by default. They're entertaining to play with, but eventually the novelty wears off and you realize they're just not good enough. Esp for tasks where correctness is important (summarization, for example).